欢迎访问《人工晶体学报》官方网站,今天是 分享到:

人工晶体学报 ›› 2026, Vol. 55 ›› Issue (3): 368-377.DOI: 10.16553/j.cnki.issn1000-985x.2025.0178

• 研究论文 • 上一篇    下一篇

改进的YOLO11晶体表面缺陷光学检测方法

殷创业1,2(), 王华东2, 张庆礼2, 孙贵华2, 孙彧2, 张志荣1,2()   

  1. 1.合肥大学电子信息与自动化学院,合肥 230601
    2.中国科学院合肥物质科学研究院安徽光学精密机械研究所,光子器件与材料安徽省重点实验室,合肥 230031
  • 收稿日期:2025-08-06 出版日期:2026-03-20 发布日期:2026-04-08
  • 通信作者: 张志荣,博士,研究员。E-mail:zhangzr@aiofm.ac.cn
  • 作者简介:殷创业(1998—),男,安徽省人,硕士研究生。E-mail:17856539063@163.com
  • 基金资助:
    先进激光技术安徽省实验室开放基金主任基金(AHL2021ZR05);中国科学院合肥大科学中心协同创新培育基金重点项目(2021HSC-CIP005)

Improved YOLO11 Optical Detection Method for Crystalline Surface Defects

YIN Chuangye1,2(), WANG Huadong2, ZHANG Qingli2, SUN Guihua2, SUN Yu2, ZHANG Zhirong1,2()   

  1. 1.School of Electronic Information and Automation,Hefei University,Hefei 230601,China
    2.Anhui Provincial Key Laboratory of Photonic Devices and Materials,Anhui Institute of Optics and Precision Machinery,Hefei Institute of Physical Sciences,Chinese Academy of Sciences,Hefei 230031,China
  • Received:2025-08-06 Online:2026-03-20 Published:2026-04-08

摘要: 晶体制品在生产、制作、加工过程中易产生划痕缺陷,精准识别这些缺陷成为该领域的技术难题。为应对复杂成像条件下的多类型划痕检测的要求,本文基于YOLO11,引入可变形的注意力转换器(DAT)和大可分离核注意力(LSKA)机制,通过采用可变性卷积增强图像的特征输出,以及引入不同尺度感受野机制,增强模型对晶体表面多尺度缺陷特征的自适应建模能力。相较于传统的 YOLOv3、YOLOv5、YOLOv8及本文的基线模型YOLO11,改进后的YOLO11-DAT_LSKA在mAP@0.5指标上分别对相关缺陷进行检测分析,取得了2.5、2.3、1.9和1.5个百分点的性能提升。说明该处理方法在增强特征建模能力和提升对复杂划痕的感知具有一定优势,提高了模型对晶体表面缺陷的检测精度。特别是针对晶体表面缺陷具有低对比度、强反射、细长不规则形态等特征,本文所引入的DAT与LSKA模块分别从可变形卷积与多尺度感受野自适应两个层面出发,使模型在刻画缺陷形态差异和多尺度缺陷结构方面具有针对性与鲁棒性。

关键词: 晶体表面缺陷; YOLO11; 可变形的注意力转换器; 大可分离核注意力; 注意力机制

Abstract: Crystal products are prone to scratch defects during production, manufacturing, and processing, and accurate identification of such defects remains a technical challenge in this field. To address the urgent demand for multi-type scratch detection under complex imaging conditions, this paper proposes an improved approach based on YOLO11 by introducing a deformable attention transformer (DAT) and large separable kernel attention (LSKA). The method improves adaptive modeling and recognition of surface defects by leveraging deformable convolutions and multi-scale receptive field fusion. Compared with traditional models such as YOLOv3, YOLOv5, YOLOv8, and the baseline YOLO11, the improved YOLO11-DAT_LSKA achieves performance improvements of 2.5, 2.3, 1.9 and 1.5 percentage points respectively in terms of mAP@0.5 for defect detection. These results demonstrate the effectiveness of the proposed method in enhancing feature modeling capabilities and improving the perception of complex scratches, thereby improving the detection accuracy for surface defects on crystal materials. In particular, to address low-contrast and irregular crystal surface defects, the DAT and LSKA modules enhance global contextual modeling and multi-scale receptive field adaptation, enabling effective capture of defect shape variations and multi-scale defect structures.

Key words: crystalline surface defect; YOLO11; deformable attention transformer; large separable kernel attention; attention mechanism

中图分类号: